September 2026 Summaries

Code review is often slow because humans spend time inspecting implementation details that automated agents can check more quickly and consistently, such as null handling, complexity, security patterns, and test results. The argument proposes shifting human review toward assumptions, product intent, interfaces, long-term compatibility, operational safety, and rollout plans, since these judgments require contextual understanding that code-focused tools cannot reliably provide. Teams should document key assumptions before implementation, identify which are verified or unverified, and have reviewers challenge the highest-risk premises early, before completed code makes fundamental changes harder to raise. Pull request reviews should focus on user impact, API and configuration commitments, reversibility, monitoring, and newly discovered assumptions, while human line-by-line inspection is reserved for explicitly sensitive areas such as tenant isolation, authentication, billing, and deletion. Finally, teams should validate important assumptions after release through attached metrics, creating a feedback loop that improves future decisions and relocates review effort from diffs to the questions machines cannot answer.

Redocly’s MCP server now supports code mode, which allows AI agents to write and execute JavaScript in a server-side sandbox to retrieve, filter, and analyze API documentation before returning only relevant findings to the model. Unlike direct tool calling, where each intermediate tool result enters the LLM context, code mode can handle multi-step tasks such as identifying API operations without documented 400 responses through loops and conditional logic, reducing unnecessary context use and LLM decisions. In a sample audit of three APIs, Redocly reported 91% fewer client-visible calls, 80% fewer input tokens, and 79% lower cost while maintaining equivalent or better answer quality. Across multiple documentation tasks and five models, the company measured input-token savings of 50–90% and cost reductions of roughly 45–80%, especially for queries requiring broad retrieval, comparison, or filtering. Code mode is available to Redocly MCP server users and is also used by its AI Assistant, with future plans to enable agents to call users’ APIs as well.

Redocly’s experimental `generate-spec` CLI command creates an initial OpenAPI 3.2 description from recorded HTTP traffic, addressing APIs that lack specifications needed for documentation, SDKs, and AI-agent integrations. It accepts HAR files and common proxy or server logs, deterministically inferring templated paths, request and response schemas, optional fields, repeated object components, selected formats, and cautiously identified enums from observed exchanges. An optional AI refinement mode processes one operation at a time using limited real traffic samples, validates proposed changes against the baseline, and can add descriptions, examples, constraints, semantic types, and more expressive structures such as discriminated `oneOf` payloads. In a Cafe API evaluation against a handwritten specification, AI modestly improved response-schema recovery and substantially enhanced documentation and request-body typing, including correcting multipart form values that traffic alone represented as strings. The tool runs through locally installed Claude, Codex, or Cursor CLIs, supports parallel processing, and omits headers, isolates execution environments, and instructs models to scrub secrets, though users are advised to capture only clean synthetic test traffic because request data is sent to external AI providers. Generated specifications remain hypotheses limited by observed coverage, may overfit sampled values, cannot reliably recognize ordinary-word path parameters, and require human review and ongoing drift checks.

The passage argues that AI agents make parallel work possible without the attention-splitting costs associated with human multitasking, because each agent independently handles a single task while the user manages a queue of assignments. It distinguishes multitasking, in which one person repeatedly shifts focus between tasks, from parallelism, in which multiple workers pursue separate tasks simultaneously. Using an archer metaphor, it emphasizes that the most important work occurs before an agent begins: users should clearly define the target, success criteria, risks, and stopping condition in a concise prompt or “pitch,” then let the agent execute without continual intervention. Anxiety about running several agents at once is presented as a sign of poorly scoped or insufficiently planned tasks rather than a flaw in parallelism itself. The proposed skill for an agent-enabled workplace is to repeatedly prepare, launch, and review well-defined tasks while other agents work independently.

Redocly’s experimental `introspect-mcp` CLI command helps document Model Context Protocol servers by querying a running server’s tools, prompts, resources, schemas, capabilities, and protocol version, then recording them in the `x-mcp` OpenAPI extension. It supports remote servers over Streamable HTTP or legacy HTTP+SSE, authenticated endpoints through custom headers, and locally launched MCP packages through stdio using the `--command` option. The generated or updated OpenAPI file can be rendered in Redocly Realm alongside conventional API reference documentation, making runtime-discoverable MCP capabilities visible to human users. Refreshing preserves existing OpenAPI content and selected manual annotations such as tags, security settings, and prompt examples while replacing introspected entries to reflect renamed or removed capabilities. A `--check` mode can be added to CI pipelines to detect documentation drift and fail builds when the recorded MCP description differs from the live server.

A year after initially criticizing llms.txt, the author reports that server logs, an Ahrefs analysis, and statements from Google continue to show little evidence that AI crawlers or chat systems routinely fetch or use the root-level manifest, while most observed requests come from SEO tools, scanners, or site owners themselves. Although adoption has grown through automated documentation platforms, plugins, and audit tools, the author argues that this reflects hype rather than demonstrated demand, especially since major AI providers have not said they consume others’ files. The more durable contribution of the proposal has been clean, predictable Markdown versions of documentation pages, which coding agents frequently use because they are token-efficient and avoid interface clutter; the specification itself has shifted toward helping agents locate those Markdown pages. The author’s organization still generates llms.txt at no cost but prioritizes improved Markdown, agent feedback mechanisms, MCP servers for source-controlled and access-aware documentation retrieval, and service descriptions such as MCP and A2A cards. The piece concludes that organizations seeking AI readiness or governance should focus on usable content formats, direct agent integrations, licensing and legal clarity, while treating llms.txt as a low-priority file that may also require review for potential prompt-injection risks.

Redoc CE 3 is an unreleased major overhaul of the open-source API documentation renderer, designed to address Redoc 2’s aging interface, limited theming, poor large-specification performance, and technical debt. After initially publishing an OpenAPI-only release candidate in late 2025, the team restarted development in January 2026 to build a shared architecture that parses specifications into structured data before rendering, enabling support for OpenAPI 3.2, AsyncAPI, GraphQL, and MCP definitions through OpenAPI’s x-mcp extension. The new version adds CSS-custom-property theming, built-in dark mode, viewport-aware rendering, expanded search, per-entity routes and precise deep links, while reducing configuration options and replacing the JavaScript theme object. It will use ESM bundles and require React 19 and Node 22, though Redoc 2 will continue receiving security and critical fixes. Redoc CE 3 will be available through HTML, React, CLI, Docker, and cloud-hosted deployments, sharing its core codebase with Redocly’s commercial products while retaining commercial-only interactive features such as Try It. It will also introduce opt-out anonymous OpenTelemetry-based usage telemetry, and users are encouraged to test the demo and provide feedback before rc.1 and the eventual stable 3.0 release.

Redocly CLI’s `generate-arazzo` command can convert OpenAPI descriptions into Arazzo workflow files, with its default mode producing fast, deterministic single-operation scaffolds that require users to manually connect dependencies between API calls. The new `--with-ai` option instead uses a locally installed Claude Code, Codex CLI, or Cursor CLI assistant to infer realistic multi-step API scenarios, such as registering an OAuth2 client, creating a resource, retrieving it, and deleting it while passing generated values between steps. AI-generated workflows include schema-compliant payloads, security configuration, inferred authentication flows, and linked outputs and parameters, but they are validated against the source API description and fall back to the standard skeleton if generation fails or is invalid. Redocly does not receive API keys or workflow data, though the resolved OpenAPI description is sent to the chosen AI provider, so users are advised to remove sensitive information. Options allow selection of provider and model, workflow limits, and concurrency, while large specifications are processed through a two-phase scenario-selection approach; generated workflows should be reviewed because they vary between runs.

Redocly has introduced the experimental open-source `generate-client` command in its CLI and the accompanying `@redocly/client-generator` package, which generate typed SDKs from OpenAPI 3.0–3.2 or Swagger 2.0 definitions for TypeScript, Python, Go, and PHP. A single specification can also produce validation schemas, TanStack Query and SWR integrations, deterministic test mocks, command-line interfaces, data transformers, and regenerable reference documentation. Generated clients include features commonly written manually, such as scoped authentication, pagination iterators, retries, middleware, query serialization, multipart uploads, typed Server-Sent Events, error-handling options, and strict compile-time or runtime argument validation. The tool is designed to support AI-agent workflows by making API contracts discoverable through generated CLI help and schemas, providing instruction “skills” for safely modifying generators, and encouraging regeneration instead of editing generated HTTP code. Organizations can eject built-in generators into their repositories for customization without fully forking the project, retain upstream updates, and define local design rules for agents to follow. Redocly reports that it has used the generator for four internal APIs and production traffic, while cautioning that the command remains experimental and users should pin CLI versions because flags and output may change.